Experiment Tracking with MLflow

Machine learning development is iterative and empirical. Without rigorous experiment tracking, models cannot be reliably reproduced, audited, or safely promoted to production staging.

Core MLflow Abstractions

• Experiments: Logical grouping of runs addressing the same problem domain (e.g., churn-prediction-v2).
• Runs: A single execution of machine learning code recording:
  • Parameters: Learning rate, batch size, optimizer choice (mlflow.log_param).
  • Metrics: Training loss, validation accuracy, F1 score recorded across epochs (mlflow.log_metric).
  • Artifacts: Serialized model weights (model.pt), ONNX files, confusion matrices, requirements.txt.
  • Tags: Git commit hash, training dataset version, developer ID.

Model Registry Lifecycle

Once a run achieves target performance, its weights are registered in the MLflow Model Registry: Draft $\rightarrow$ Candidate $\rightarrow$ Staging $\rightarrow$ Production.

What you'll practice

• Register the Champion Model — practice · intermediate

A command you'll meet

echo stage=Production > stage.txt

A registry is a directory tree plus discipline: one folder per model, one per version, and a stage label that says out loud what is serving real traffic. Open the task below for the full walkthrough.

What you'll practice

• Tag Every Model Version — practice · intermediate
• Promote Models with Aliases — challenge · advanced

A command you'll meet

cat ~/mlflow/tags.yaml

Six months later, the tags are the only thing that explains the model. Open the task below for the full walkthrough.